Translate and deliver to Eunji

Step 1: Complete the methodology memo

Open materials/methodology-memo-template.md and fill in every remaining section. The Preparation Decisions section should now be complete: feature engineering rationale, temporal split justification, leakage assessment with the comparison results, and data quality issues (stockouts, platform migration).

Fill in the Methodology section (model choice, evaluation metrics), the Limitations section (what the model cannot predict, where forecast reliability breaks down), and the Findings Summary.

The methodology memo documents your preparation pipeline. If another analyst inherited this project and needed to understand why you lagged the social media features by seven days instead of using same-day counts, this memo should answer that question without them needing to ask you.

Step 2: Cross-model review

Open a second Claude session. Give the second AI only three things:

1. Eunji's original Slack message
2. The data dictionary
3. Your completed methodology memo

Ask the review model: "Does the feature engineering avoid leakage? Is the temporal split correct? Are the limitations honest? Did I handle the stockout data appropriately?"

Read the review. It may catch things you missed -- a feature that is partially leaked, a limitation you did not state explicitly, or a preparation decision that needs more justification. Address any legitimate findings.

Step 3: Translate the forecast into buying language

The buying team does not read MAE numbers. They need to know: what should we order, how confident are we, and what should we watch.

Translate the forecast into operational language:

• Not "MAE = 47 units" but "for seasonal products, the forecast is usually within 50 units of actual demand -- close enough for weekly ordering"
• Not "trend-driven products have higher MAE" but "for products that go viral, the model catches trends that are building on social media about a week ahead, but it cannot predict products that spike with no warning"
• Not "feature importances show calendar variables dominate" but "for sunscreen and moisturizer, past years' patterns tell us what to order -- social media does not add much for these products"

Step 4: Address Eunji's five requirements

Go back to her original Slack message. She asked for five things:

1. Predict demand for the top 200 SKUs at least a week ahead
2. Flag products that are starting to trend upward before they spike
3. Separate seasonal products from trend-driven products
4. Tell her what data signals are actually predictive
5. Something the buying team can read and act on

Map each finding to one of these requirements. Every requirement should have a clear response in the findings summary. If a requirement is only partially met (and some will be -- trend prediction is inherently limited), say so.

Step 5: Draft the findings summary

Structure the findings summary around what the buying team needs to do, not around the methodology. Lead with actionable information: which SKUs to order more of, which products are trending upward, which seasonal patterns are coming. Put limitations alongside the relevant findings, not in a separate section at the bottom.

The summary is the deliverable. The notebook and methodology memo are supporting evidence. Eunji's buying team will read the summary. They will not read the notebook.